add server zero overhead engine

vllm/engine/async_llm_engine.py

@@ -280,6 +280,8 @@ class _AsyncLLMEngine(LLMEngine):
         """
         # these are cached outputs from previous iterations. None if on first
         # iteration
+        if self.zero_overhead:
+            return self.zero_overhead_step()
         cached_outputs = self.cached_scheduler_outputs[virtual_engine]
         seq_group_metadata_list = cached_outputs.seq_group_metadata_list
         scheduler_outputs = cached_outputs.scheduler_outputs

vllm/worker/model_runner.py

@@ -60,7 +60,7 @@ from vllm.worker.model_runner_base import (
     _init_attn_metadata_from_tensor_dict,
     _init_sampling_metadata_from_tensor_dict)
 
-from vllm.model_executor.layers.ops.update_input import UpdateInputTokens
+from vllm.model_executor.layers.update_input import UpdateInputTokens
 
 if TYPE_CHECKING:
     from vllm.attention.backends.abstract import AttentionBackend